Blog

Last month I wrote that ignorance is an advantage. This month agents burned me four times: performance inflated by orders of magnitude, a lab machine bricked, optimizations going in circles, and goals too high to ship. Ignorance gets you started; common sense keeps you from getting played.

Today I went to the AMD Developer Conference in Shanghai. Over 2,000 people packed the venue, and AMD gave one of its two developer awards to a non-coder who used an AI agent to rewrite a project in Rust. A year ago everyone worried about demand; six months later, all anyone asks is: 'Got more supply?'

When I show friends what my AI agents are doing, they always ask, 'Can you explain this to my boss?' But real change has never come from a single lesson.

I recently worked on a few AI content batching projects and realized something I'd miscalculated: going from a working demo to a fully automated pipeline takes at least 3× the time, not 50%. The extra two hours go entirely into QA and retries. Here's what I learned about where mid-tier models and local compute actually fit.

I spent two days automating with the Codex app, burned through a Pro account, and made virtually no progress. Switching to Codex CLI's Goal feature fixed everything immediately. At first I thought the model had gotten dumber, but then I realized: the agent form factor is solidifying, and humans shouldn't be chained to the screen anymore.

I ran seven hundred rounds of AI Infra experiments, and thirty-five hours were entirely eaten up by environment startup. At first I thought GPT-5.5 fast mode wasn't fast enough, but later realized it wasn't the model thinking—it was the model waiting for the CPU. Intel has already tightened the server CPU:GPU ratio from 1:8 to 1:1.

Claude told me to 'get some sleep' halfway through a task, and GPT-5.5 killed the direction after a quick probe. After a week with the new models, one thing became clear: no matter how powerful the model gets, the collaborative part still falls on humans.

A few decades ago, AI meant machine learning. Three years ago, it was ChatGPT. This year, it's agents. The word never changed, but what it contains has been replaced three times over. Most people are still stuck in the previous generation's cognitive framework, discussing something that doesn't actually exist.

A tool I couldn't build in two days was done in half an hour by switching models. When I couldn't log into my computer, an AI agent fixed it in half an hour. Later, I built Aima Service—so every device can have an AI ready to help when things go wrong.

Zero-friction onboarding plus extremely powerful AI intelligence—two pillars can't hold up a product. When users see an agent executing incomprehensible commands in the terminal, their first reaction isn't awe, it's fear. The missing pillar is progressive trust.

Prompts, RAG, fine-tuning, knowledge graphs, context engineering—five paradigm shifts in three years. Models are getting stronger, but we still don't know how to make agents perform well in products.

Mac's permission model is designed for humans; Linux is designed for programs. An Agent is not an app—it is infrastructure. Running an Agent on a Mac is like stuffing a server into a laptop chassis—it works, but everything feels off.